RLGS-YOLO: an improved algorithm for metro station passenger detection based on YOLOv8

Abstract

Passenger detection is a key component of guaranteeing the safe operation of the subway. Nevertheless, the issue of varying target sizes across subway scenarios impedes passenger detection. Additionally, there is the issue of occlusion overlap between passengers. To resolve these concerns, an improved model based on YOLOv8n is proposed. It consists of Reduced Channel Spatial Object Attention (RCSOSA) module, Large Separable Kernel Attention (LSK) module, Group Shuffle Convolution (GSConv) convolution, and other modules such as Small Detect Head (S-detect), and named RLGS-YOLO. Designed to facilitate the deployment of a variety of platforms, S-detect lightweight detection head module minimizes the number of parameters and the amount of model computation. A re-parameterized structure based on Re-param Visual Geometry Group (RepVGG) is proposed to address the issue of mutual occlusion between passengers. This structure is achieved by integrating RCSOSA module, which substantially improves the information exchange between various channels. The dynamic sensory field module of LSK has been incorporated. It improved the issue of model misdetection by whole-heartedly incorporating the background information. The model's sophistication is de-creased by integrating GSConv lightweight convolution. The experimental findings indicate that in comparison to YOLOv8n. In the passenger data set of Nanning Metro Line 1, RLGS-YOLO indicates superior performance. The Mean Average Precision (mAP) improved by 2.2%. RLGS-YOLO, on the other hand, attains superior performance on the public dataset VOC2007. This contains a 1.1% increase in mAP, a 13.79% increase in Frames Per Second (FPS), and a 0.5G Floating Point Operations (FLOP) reduction in computation. In comparison to other prominent detection models, RLGS-YOLO also exhibits superior performance. The improved RLGS-YOLO model provides a precise and efficient solution for passenger detection methods.